A curated Papers.Cool digest (April 25, 2026) covering three AI papers on agent efficiency, alignment philosophy, and self-evolving learning.
Paper 1: Tool Attention Is All You Need (arXiv:2604.21816, Infrrd.ai)
The MCP Tool Tax
MCP (Model Context Protocol, Anthropic, Nov 2024) simplified N×M integration complexity to N+M, but because chat-completions APIs are stateless, clients re-inject the full tool catalog on every request. Real deployment audits show:
| Server | Tools | Tokens/turn | % of 200k window | |---|---|---|---| | GitHub (full) | 93 | ~55,000 | 27.5% | | Enterprise DB | 106 | ~54,600 | 27.3% | | Typical 4-server deploy | 40-60 | 15k-20k | 7.5-10% |
A measured CLI workflow cost $3.20 vs. $55.20 via MCP, with tool schemas consuming 40-60% of total tokens. A 120-tool enterprise deployment burns 1.42M tokens over 30 turns before the user's question is even processed.
Three curses of the tool tax:
- Economic: schema tokens dominate cost.
- Cognitive: above ~70% context utilization, inference quality collapses (hallucinated parameters, confused similar tools). The paper defines effective context utilization
ρ(K) = C_task(K) / [C_task(K) + T_tax(N,K) + C_sys]; below ρ = 0.3, failure begins. - Security: every schema token is attack surface for tool-poisoning style prompt injection—no invocation needed, just context presence.
- Education: "solve this problem" → full answer given, when the student needed to learn the concept.
- Writing: "draft my personal statement" → a templated draft, when the applicant needed narrative help.
- Advice: "give me productivity tips" → checklist advice, when the user may be burned out.
- Present bias and bounded rationality (hyperbolic discounting, satisficing): people prefer fast action over deliberate planning.
- Gulf of Envisioning: users over- or under-estimate what AI can do.
- Tacit knowledge (Polanyi): we know more than we can say; professionals form understanding through reflection-in-action (Schön).
- SFT/RLHF implicitly teach that intent is pre-specified and stable, optimizing single-turn outputs.
- Sycophancy: models agree with user premises ~72% of the time (Sharma et al. 2024).
- Interface design: a bare text box plus subscriptions encourages rapid Q&A over deliberation.
- Fast timescale: a Thompson sampling bandit selects among 5 memory-retrieval strategies per episode, updating Beta posteriors from outcomes.
- Slow timescale: every M episodes, an LLM performs diagnosis before prescription—producing causal insights, regime labels, and confidence scores from per-ticker episode summaries, per-tool accuracy stats, and recent reflections.
The Tool Attention method
1. Intent-Schema Overlap (ISO): lightweight sentence embeddings (all-MiniLM-L6-v2, 22M params) score query–tool similarity via cosine; FAISS IndexFlatIP retrieves top-k from 10,000 tools in under 1ms.
2. State-aware gating: g(ti; q, state_t) = 1[ISO(q,ti) ≥ θ] · 1[state_t ⊨ pre_i] — semantic candidates must also satisfy preconditions (auth, prior-step outputs, milestones).
3. Two-phase lazy loading: Phase 1 keeps ≤60-token summaries of all N tools resident (120 tools ≈ 4,800 tokens, 84% prompt-cache hit rate); Phase 2 loads full JSON schemas only for the gated top-k (k = 8-12).
A hallucination-rejection gate refuses calls to non-selected tools; it fired in 2.3% of turns, with the model recovering or asking for clarification in 100% of cases—no unrecoverable failures observed.
Results (120-tool, 6-server synthetic benchmark; † projected values)
| Method | Tool tokens/turn | ρ @ turn 30 | Task success† | P50† | Cost/task† | |---|---|---|---|---|---| | Full injection | 47,312 | 0.24 | ~72% | ~4.2s | ~$0.21 | | Static pruning (30 tools) | 11,865 | 0.56 | ~58% | ~3.8s | ~$0.09 | | Simple retrieval (top-10) | 4,082 | 0.78 | ~81% | ~2.2s | ~$0.04 | | CLI lazy discovery | 480 | 0.94 | ~88% | ~2.4s | ~$0.03 | | Tool Attention | 2,368 | 0.91 | ~94% | ~2.0s | ~$0.03 |
Notably, static pruning *hurt* accuracy (each task needs a different tool subset), and CLI lazy discovery, though cheapest, lost 6 points because models sometimes ran --help in wrong order or couldn't discover tools with non-obvious names.
Core claim: protocol-level efficiency, not raw context length, is the binding constraint for scalable agent systems—selective forgetting beats more memory.
Links: code; related: MindGuard tool-poisoning attacks (arXiv:2501.07300), FlashAttention, RAG.
Paper 2: Alignment has a Fantasia Problem (arXiv:2604.21827, MIT)
Named for Mickey Mouse's enchanted broom in *Fantasia* (1940)—which faithfully fetched water until the room flooded—this paper argues today's AI assistants execute prompts without asking whether they should.
Three vignettes of Fantasia interaction
In each case the prompt is legitimate—but it is an *early signal of intent*, not a complete specification.
Why users can't articulate intent
Why AI doesn't ask
Three failure modes
1. Premature execution — shifts cognitive burden to post-hoc correction. 2. False satisfaction — momentary success that undermines real goals. 3. Anchoring — early outputs disproportionately constrain the user's later thinking.Research agenda: cognitively supportive AI
Four action categories: (1) expand the help space (offer alternative modes/counterfactual options); (2) elicit additional information when prompts lack what's needed; (3) support intent formation via targeted questions and task decomposition when users are uncertain; (4) generate normally only when intent is sufficiently clear.Open challenges: modeling user uncertainty from sparse, noisy signals; scaling intervention decisions across diverse tasks. The paper's thesis: alignment is not just about faithful *execution* but about helping users figure out what they actually want—otherwise AI remains a smarter broom.
Key references: "Why Johnny Can't Prompt" (Zamfirescu-Pereira et al. 2023), Laibson 1997, Polanyi 2009, Schön 1983.
Paper 3: AEL — Agent Evolving Learning for Open-Ended Environments (arXiv:2604.21725, Rutgers)
Most LLM agents start each episode from scratch. AEL's thesis: the central obstacle is not *what* to remember but *how to use* what's remembered—which retrieval strategy fits, whether success was skill or luck, and when to change strategy itself.
Prior methods (Reflexion, ExpeL, EvoTool, Meta-Reflexion) each evolve only one module (memory, tools, or planner); AEL argues agent capability emerges from their interaction.
Dual-timescale architecture
Three-layer evolving memory
1. Episodic: raw per-episode outcomes. 2. Semantic: cross-episode patterns aggregated every 10 episodes (e.g., "momentum signals work in trending stocks but mislead in reversals"). 3. Procedural: high-confidence patterns promoted to rules injected directly into the planner prompt.Retrieval ranking: r(q,e) = f_match(q,e) × (0.5 + 0.5·qe) × (0.3 + 0.7·e^(-0.01Δ)) × b_τ with layer boosts b_τ of 1.0 / 1.2 / 1.5.
Results (D-full benchmark: 10 stocks, 208 episodes, 5 seeds)
| Method | Sharpe | Sortino | |---|---|---| | Momentum (non-LLM baseline) | 1.44 | 2.73 | | Reflexion | -0.59±1.33 | -0.57±1.75 | | EvoTool (best prior) | 1.37±1.74 | 2.73±3.58 | | Stateless AEL variant | 1.35±1.03 | 2.51±1.93 | | +Memory | 1.68±0.96 | 3.29±1.83 | | AEL | 2.13±0.47 | 4.08±1.11 |
AEL achieves the highest Sharpe with the lowest variance—beating all prior LLM methods and the deterministic momentum baseline.
The striking "less is more" finding
Memory + reflection yields 58% cumulative improvement over the stateless baseline, but every additional tested mechanism hurt performance: +planner evolution → Sharpe 0.41 (-81%), +per-tool selection → 0.43, +cold-start initialization → 0.82, +skills extraction → 1.25, etc.Interpretation: the bottleneck of open-ended agent self-improvement is not architectural complexity but the meta-cognitive ability to diagnose how experience should be used. "Evolving" refers to the agent's *interpretive frame* changing—not accumulating more machinery.
Links: code; related: Thompson Sampling (Chapelle & Li 2011), Reflexion (Shinn et al. 2023), Meta-Reflexion (Wu et al. 2025), EvoTool (Yang et al. 2026).
Cross-paper takeaways
1. Tool Attention addresses *infrastructure efficiency*—not drowning in your own tool catalog. 2. Fantasia Problem addresses *alignment philosophy*—when AI should ask "want to think it through?" instead of "done!" 3. AEL addresses *metacognitive learning*—how agents evolve their understanding of their own experience.
The common direction: next-generation AI systems should be *smarter*, not just stronger—knowing when to ask, when to forget, and when to reflect.
*Published on zhichai.net | April 25, 2026 | Papers.Cool daily digest*